Mid-scale science

Given the potential flexibility of their funding, philanthropists and foundations can work with the research community to not only identify high-impact midscale projects, but to co-design the right mechanisms to fund, organize and incentivize the R&D

There are many types of mid-scale science projects. In some cases, these might make sense to create and sustain as “public goods,” such as open datasets, or broadly-available user facilities. In other instances, the research may lead to a startup or a new commercial product or service. Although these are described below as distinct types of projects, many initiatives may involve combinations of these approaches.

1.Datasets and benchmarks.

Open-access datasets and benchmarks are allowing researchers to train powerful AI models. For example, the Protein Data Bank has enabled Nobel Prize-winning advances in both protein structure prediction (AlphaFold), and protein design (RFDiffusion). EvE Bio’s “pharmome” is mapping the unintended targets of small-molecule pharmaceutical drugs. This open dataset could help researchers predict the negative side-effects of drugs before they reach patients, and identify opportunities to repurpose existing drugs to treat a broader range of diseases. Align to Innovate is working with the research community to define and create datasets improving “structure to function” prediction using AI. You can read more about the importance of datasets and benchmarks in our playbook on the Common Task Method.

2. New platform technologies for both imaging/characterization/measurement and synthesis/fabrication/perturbation.

In many instances, our ability to understand some complex phenomena (e.g., the architecture of the human brain with respect to memory, perception, problem-solving) is limited by existing research tools. Investing in the development of new and improved tools, or lowering the cost of existing tools, can have a transformational impact on a field. Examples include a reduction in the cost of sequencing the human genome from $100 million to $100, or the impact that electron microscopes and the ability to see individual atoms has had on materials science.

3. New foundation models for science.

Although Large Language Models have been trained on text, researchers are beginning to train foundation models on scientific data. For example, Tatta Bio is developing genomic language models that are trained on trillions of base pairs from metagenomic datasets, and that can predict novel types of protein-protein interaction. It may be difficult for academic researchers to train these foundation models in the absence of philanthropic support if they require large datasets, expensive GPU clusters, and professional ML and software engineers.

4. Automation for science.

Although not all experiments can be automated, scientific automation has a number of potential benefits. It can allow researchers (particularly graduate students and postdocs) to spend more time on the design of experiments and the interpretation of results, as opposed to repetitive manual tasks. Automation can also increase the reproducibility of research results, and make it easier for researchers to incrementally increase the size of a dataset if it is particularly valuable for training an AI model. “Cloud labs” (remote access to automated scientific equipment and reagents) can expand the number of experiments that any single researcher can take advantage of. “Self-driving labs” (integrated combinations of AI, automated equipment, software for the orchestration of scientific workflow),can accelerate the pace of scientific discovery. AI can create a system of closed loop experimentation by identifying the most valuable experiment to do next. In many fields, human judgment and intuition are still playing an important role, so developers of self-driving labs are working on designing the right forms of human-machine interaction.

5. Modern scientific software.

Although university researchers develop academic prototypes of scientific software, they often lack the incentive, talent and funding to develop and maintain “production-quality” code. In many instances, researchers also lack the resources to rewrite legacy code so that it is written in a modern language, optimized for modern computer architectures such as GPUs, and is designed to be integrated with AI. For example, some researchers are producing scientific software so that key parameters in a model can be “learned” and therefore improved over time.

6. Foundries or shared facilities.

Some resources needed for scientific research are expensive, or also require access to specialized expertise. For example, in the 1980s, DARPA accelerated progress in the field of microelectronics with a program called MOSIS, which gave academics and small businesses access to the ability to prototype new chip designs.

7. Sector-specific public goods.

One example is the Fusion Prototypic Neutron Source (FPNS), which will allow research to discover radiation tolerant materials for fusion reactors. Investment in an FPNS would benefit many fusion companies, in the same way that federal investment in wind tunnels in the 1930s strengthened U.S. leadership in aeronautics.

FROs
An example of mid-scale science in practice is the Focused Research Organization (FRO) model. FROs build a startup-inspired, non-academic team and a transformative goal with defined deliverables rather than merely a theme or set of questions. Convergent Research is a non-profit studio and parent organization for FROs and defines them as follows:

Focused
FROs pursue pre-specified, quantifiable technical milestones rather than open-ended, basic research. Teams are funded to aggressively pursue these milestones within a finite time – usually 5 to 7 years – to avoid mission creep and preserve focus.

Research
FROs undertake research with technical risk to produce high-impact public goods in key areas of science and technology. They can result in products like massive datasets, next-generation analytical devices, and open-source experimental protocols. FRO projects are technically ambitious and often engineering-heavy.

Organizations
FROs are led by a full-time founding team, and typically consist of 10 to 30 scientists, engineers, and operational staff. FROs execute like the best deep-tech startups: tight-knit, fast-moving, and mission-driven.